KEGG orthology-based machine learning reveals functional determinants of antimicrobial resistance in Acinetobacter baumannii

ABSTRACT Antimicrobial resistance represents a critical global health threat, with Acinetobacter baumannii classified as a priority pathogen due to its extensive multidrug resistance. Currently, antimicrobial susceptibility testing relies on phenotypic characterization, which provides limited mechanistic insights into resistance mechanisms and significantly delays appropriate therapy. While machine learning approaches have shown promise for rapid resistance prediction, existing computational methods rely predominantly on gene presence-absence matrices that inadequately capture functional capacity. To address this limitation, we developed a functional genomics framework using KEGG orthology annotations to predict antimicrobial resistance phenotypes directly from whole-genome sequencing data. We systematically evaluated six machine learning algorithms on 1,088 A. baumannii isolates across 10 antimicrobial agents, with subsequent validation in an independent cohort of 508 isolates. KEGG orthology-based representations consistently outperformed traditional gene presence-absence encoding across all algorithms. XGBoost achieved superior performance (94.44% weighted accuracy) and was therefore selected for interpretability analysis. Identified determinants included known carbapenemases and, crucially, K18974 ( sul1 ), a ubiquitous marker of integron-mediated propagation. Additionally, K21801 ( iaaH ) was identified as a key driver of non-conventional resistance via indole-mediated signaling. Subsequent multidrug resistance analysis identified shared determinants—overlapping KEGG orthologs across antimicrobial classes—that generate predictable cross-resistance patterns. These computationally derived molecular networks closely mirrored the observed clinical co-resistance phenotypes. Overall, this interpretable machine learning framework advances mechanistic understanding of resistance evolution and establishes a promising computational foundation. Future prospective validation is required to translate these findings into clinical utility. IMPORTANCE Acinetobacter baumannii represents one of the most urgent antimicrobial resistance threats globally, and traditional diagnostic methods can take up to 48 hours to determine appropriate treatment options. This delay contributes to poor patient outcomes and drives further resistance development through inappropriate antibiotic use. This study addresses a critical gap by developing a machine learning framework that can rapidly predict antimicrobial resistance patterns directly from whole-genome sequencing data. Unlike previous approaches that rely on simple gene presence-absence data, our method uses functional genomic annotations to capture the complex biological processes underlying resistance. The framework not only achieves superior prediction accuracy but also provides interpretable insights into resistance mechanisms, identifying both known resistance genes and previously unrecognized metabolic pathways involved in antimicrobial resistance. The biological insights gained can open new avenues for developing targeted therapeutic interventions.